pluckr

---

Why

Today, LLM agents only have three options for docs:

• dump entire sites into context — wasteful, expensive, fragile
• browse live — slow, flaky, blocked by SPAs and rate limits
• hope training data is fresh — it isn't

pluckr is option four. One static binary keeps a folder of clean markdown plus a tiny SQLite full-text index. An MCP server exposes that cache to any compatible host (Claude Code, Cursor, Claude Desktop, Cline, Zed, …). Your agent searches once, gets exactly the section it needs, and cites the original URL.

See it in action

$ pluckr add https://react.dev/reference --pull
  added react.dev (website) → https://react.dev/reference
  pulling react.dev (website)...
    react.dev: 412 pages, 3,180 chunks, 6 skipped, 0 errors in 18.2s

$ pluckr search "useEffect cleanup"
1. react.dev › Reference › useEffect › Cleaning up an effect
   https://react.dev/reference/react/useEffect#cleaning-up-an-effect
   ... return a [cleanup] function from the Effect ...

$ pluckr mcp     # serve the cache over MCP/stdio

Install

pluckr ships as a single static Go binary. Three install paths today:

# 1. Latest release (recommended)
go install github.com/SarthakShrivastav-a/pluckr/cmd/pluckr@latest

# 2. Specific version
go install github.com/SarthakShrivastav-a/pluckr/cmd/[email protected]

# 3. From source
git clone https://github.com/SarthakShrivastav-a/pluckr.git
cd pluckr && go build -o pluckr ./cmd/pluckr

Requires Go 1.25+. Pre-built binaries for Linux / macOS / Windows on amd64 and arm64 are attached to every GitHub release — download, unzip, drop on PATH.

Quick start

# subscribe to a few sources (kind is detected from the spec)
pluckr add https://react.dev/reference
pluckr add https://docs.python.org/3 --max 200
pluckr add facebook/react/docs              # github
pluckr add https://example.com/llms.txt     # llms.txt convention
pluckr add ~/internal-docs                  # local markdown folder

# fetch + index everything
pluckr pull --all

# search from the CLI
pluckr search "useState"

# serve to your agent
pluckr mcp

Source kinds

Kind	Spec	What it does
website	https://react.dev/reference	Sitemap → nav → BFS crawl, fetch each URL, render to clean markdown
llms_txt	https://example.com/llms.txt	Prefer /llms-full.txt if present; otherwise parse links from /llms.txt
github	facebook/react/docs	List the repo tree, pull every .md / .markdown / .mdx under the optional subdir
local	~/internal-docs	Walk a folder, pick up .md / .markdown / .mdx / .txt. No network

The kind is detected from the spec; pass --kind to override.

Hooking it up to an agent

Claude Code (recommended)

Install the bundled plugin — MCP server + skill + slash commands + a SessionStart hook in one shot:

/plugin marketplace add SarthakShrivastav-a/pluckr
/plugin install pluckr@pluckr
/reload-plugins

What you get:

• MCP server with seven tools (search/get/list/outline + refresh/add/remove)
• Skill that tells Claude when to use the cache before reaching for the web
• Slash commands — /pluckr-add, /pluckr-list, /pluckr-search, /pluckr-refresh
• SessionStart hook — runs pluckr list so Claude knows what sources exist from message zero

Or manually wire just the MCP server:

claude mcp add pluckr -- pluckr mcp

Cursor

Add to ~/.cursor/mcp.json:

{ "mcpServers": { "pluckr": { "command": "pluckr", "args": ["mcp"] } } }

Claude Desktop

Add to claude_desktop_config.json:

{ "mcpServers": { "pluckr": { "command": "pluckr", "args": ["mcp"] } } }

MCP tool surface